Multi-State Recognition of Electro-Hydraulic Servo Fatigue Testers via Spatiotemporal Fusion and Bidirectional Cross-Attention.

Electro-hydraulic servo fatigue testing machines are susceptible to concurrent degradation and failure of multiple components during high-frequency, high-load, and long-duration cyclic operations, posing significant challenges for online health monitoring. To address this, this paper proposes a multi-state recognition method based on spatiotemporal feature fusion and bidirectional cross-attention. The method employs a Bidirectional Temporal Convolutional Network (BiTCN) to extract multi-scale local features, a Bidirectional Gated Recurrent Unit (BiGRU) to capture forward and backward temporal dependencies, and Bidirectional Cross-Attention (BiCrossAttention) to achieve fine-grained bidirectional interaction and fusion of spatial and temporal features. During training, GradNorm is introduced to dynamically balance task weights and mitigate gradient conflicts. Experimental validation was conducted using a real-world multi-sensor dataset collected from an SDZ0100 electro-hydraulic servo fatigue testing machine. The results show that on the validation set, the cooler and servo valve achieved both accuracy and F1-scores of 100%, the motor-pump unit achieved an accuracy of 98.32% and an F1-score of 97.72%, and the servo actuator achieved an accuracy of 96.39% and an F1-score of 95.83%. Compared to single-task models with the same backbone, multi-task learning improved performance by approximately 3% to 4% for the hydraulic pump and servo actuator tasks, while significantly reducing overall deployment resources. Compared to single-task baselines, multi-task learning improves performance by 3-4% while reducing deployment parameters by 75%. Ablation studies further confirmed the critical contributions of the bidirectional structure and individual components, as well as the effectiveness of GradNorm in multi-task learning for testing machines, achieving an average F1-score of 98.38%. The method also demonstrated strong robustness under varying learning rates and resampling conditions. Compared to various deep learning and fusion baseline methods, the proposed approach achieved optimal performance in most tasks. This study provides an effective technical solution for high-precision, lightweight, and robust online health monitoring of electro-hydraulic servo fatigue testing machines under complex operating conditions.